Stateful synthetic dialogue for conversational AI evaluation and research.
Rhemix is a Python ecosystem for building synthetic dialogue datasets from domain-specific ingredients. A domain pack defines the world: participant roles, prompts, tools, constraints, and quality checks. Profile generation creates concrete participants for that world. The engine then runs multi-party conversations through explicit discourse-state operations, producing transcripts, state snapshots, tool traces, and generation metadata.
Rhemix is built for evaluation engineers, data scientists, and researchers who need controlled dialogue data rather than scraped, hand-authored, or single-pass conversations.
It is designed for work such as conversational AI evaluation, redaction and compliance testing, dialogue-system regression tests, and research workflows that need structured conversational state alongside plain text.
The framework separates conversation generation from domain content.
| Package | Role |
|---|---|
rhemix-engine |
Domain-agnostic runtime for discourse state, orchestration, dashboard workflows, and batch generation. |
rhemix-profiles |
Infrastructure for generating and managing participant profiles. |
| Domain packs | Vertical-specific roles, schemas, prompts, tools, constraints, routing policies, and quality checks. |
Domain packs can vary the participants, world state, scenario parameters, channel, language configuration, output style, tool surface, and evaluation rules without changing the core runtime.
Rhemix conversations are generated as inspectable artifacts, not just text transcripts.
- Conversation transcripts
- Discourse-state snapshots
- Tool traces
- Participant/profile metadata
- Generation metadata
- Domain-specific quality signals
The name is a portmanteau of rheme and remix. A rheme is the new-information contribution an utterance makes to a discourse. Remix points to the system's method: participant profiles, scenarios, domain constraints, and discourse operations are recombined to generate new conversations with inspectable structure.
Public repositories and documentation are being prepared. The public website is available at rhemix.ai.
For early access or questions, reach out at jake@rhemix.ai.
